Ant Colony Clustering Applied to Electricity Customer Grouping

This Doctoral thesis presents EPACC (Electrical Pattern Ant Colony Clustering), an original algorithm developed to group electrical load patterns on the basis of their shape. The EPACC algorithm is a clustering technique which takes as input the number of clusters and the initial centroid model composed of the less correlated patterns. The initial set of centroids is a guideline for the evolution of the clustering algorithm, with centroids evolving during the iterative process until the stabilization. So, it is necessary to check at the end of the algorithm that the centroids remain in the same position for a given number of iterations. In terms of benchmark, the algorithm was compared with the classical k-means, because both are based on an initial centroid model. It is proven that the results are better than the ones given by the k-means in the application to electrical load pattern clustering. EPACC is based on ant colony concepts and is a multi-iteration multiagent process. In ant colony clustering, the pheromone matrix expresses the link between the patterns (corresponding to the matrix rows) and the clusters (corresponding to the matrix columns). The algorithm uses the pheromone matrix as adaptive memory (contains the information on the history of the clustering process) and updates at each iteration this matrix. At each iteration, each ant proposes a solution with RLP partitioning into clusters and calculates the corresponding centroids. Each solution is evaluated by a tness indicator expressing its performance. At the end of each iteration, the pheromone matrix components referring to the best solution are reinforced. The initialization and the rst iteration are run only once, while the successive iterations are run until the stop criterion is satisfied. The stop criterion is tracking the highest tness during the iterative process and the persistence of the centroids in the same order during the iterations. The iterative process stops if the centroids remain stable after a given number of successive iterations.